What latency means on a voice call
Latency is the silence between the caller finishing a sentence and the agent starting its reply. It is the single number that most decides whether an AI call feels like talking to a person or to a machine. Vendors quote it in different ways — model speed, time to first token, voice start time — but the only one a caller experiences is end to end: they stop, and how long until they hear a voice.
Human conversation is fast. A 2009 study in PNAS led by Tanya Stivers, across ten languages, found typical gaps between speaking turns of around 200 milliseconds. AI voice agents cannot match that yet, because they have to recognise speech, think and generate a voice in sequence. The aim is to get close enough that callers stop noticing.
The latency budget: where the time goes
| Stage | What happens | What makes it slow | How good systems cut it |
|---|---|---|---|
| Endpointing | Deciding the caller has finished, not just paused | Waiting too long to be sure; noisy lines | Smarter end-of-turn detection instead of a fixed silence timer |
| Speech recognition (ASR) | Turning the last words into text | Processing audio only after the caller stops | Streaming recognition while the caller is still talking |
| Language model (LLM) | Reading the conversation and choosing a reply or action | Large models, long prompts, slow tool calls | Right-sized models, tight prompts, starting the reply before it is complete |
| Text-to-speech (TTS) | Generating the first audio of the reply | Synthesising the whole sentence before playing it | Streaming TTS that starts speaking on the first words |
| Network and telephony | Moving audio between the phone network and servers | Servers far from India, extra hops, weak mobile signal | Infrastructure close to Indian telecom networks, fewer hops |
Each stage adds its own slice, and the slices add up. A vendor whose model is very fast but whose servers sit on another continent can still feel slow to a caller in Indore. That is why the end-to-end number is the only fair comparison.
Why 0.8 to 1.3 seconds on average matters
Below roughly a second and a half, most callers accept the pause as someone thinking. Much beyond two seconds, they start filling the silence — “hello?”, “are you there?” — and now the agent and caller talk over each other, which makes the next turn worse. On a sales call that loss of rhythm costs trust; on a reminder call it makes people hang up.
Outpero's end-to-end latency is 0.8 to 1.3 seconds on average from the caller stopping to the agent speaking. For calls where that rhythm matters most, the Premium tier at ₹6.8 per minute is the one Outpero describes as ultra-low latency, with curated native-speaker voices. Standard at ₹5 suits everyday and bulk calls.
How to measure latency with a stopwatch
You do not need special tools to compare vendors honestly. You need a phone, a second device and a few minutes.
Hear it on a real call before you read further?
- 01Have the agent call your mobile, or call its demo number, on an ordinary mobile network rather than Wi-Fi calling.
- 02Put the call on speaker and record it with a second device, or run a stopwatch app on the second device.
- 03Ask short, simple questions such as “what are your timings?” so you are measuring speed, not a long thought process.
- 04Start timing the instant you stop speaking and stop it when the agent's first sound begins.
- 05Repeat at least ten times, including once where you speak in mixed Hindi and English, and note the typical value and the slowest.
- 06For more precision, open the recording in any free audio editor and measure the silent gap on the waveform.
Do this for every vendor on the same phone, network and time of day. The slowest turns matter as much as the typical one — a single five-second silence in a call is what the caller remembers.
Why latency is harder in Indian languages
Indian callers pause in the middle of thoughts — “haan… matlab… woh flat jo aapne dikhaya tha” — and switch into English for the key noun. An agent tuned to reply quickly can cut such a caller off; one tuned to wait politely can feel slow. Good endpointing has to tell a thinking pause from a finished sentence in Hindi, Telugu or Tamil, not just in English. When you test, include these natural hesitations, because a demo script read cleanly in English will flatter every vendor.
The same applies whichever direction the call goes. On Outpero's inbound calls, where the AI employee answers your business number 24/7, the caller has chosen to ring you and expects the replies to come as briskly as a good receptionist's.
Latency is not the only thing that should be fast
- —Barge-in: when the caller interrupts, the agent should stop talking almost immediately.
- —Tool calls: booking a calendar slot or checking a spreadsheet mid-call should not create a long silence; a short “one moment” helps.
- —Speed to lead: Outpero dials a new enquiry within 30 seconds, which is a different kind of latency but often matters more for sales.
- —Language switching: moving from Telugu into English should not add a pause.
Cost and compliance
Latency does not change Outpero's pricing model: no setup fee, no monthly base fee, pay per connected minute from prepaid credits, with 50 free credits at signup to run your own stopwatch test. Compliance is unaffected too — calls go out on a DLT-registered line under TRAI's TCCCP 2018 rules, with DND screening on bulk campaigns.
Limits
No vendor controls the caller's mobile signal; a weak 4G connection on a highway will add delay no matter whose agent is speaking. Very long or complex answers also take longer to start. And a fast agent with a poor script still makes a poor call — treat latency as a requirement to pass, then judge the conversation itself.